perf(bigquery): [T1-17] eliminate unconditional HashMap allocations and dead STRUCT pass in QueryParameterValue

[lqiu96]

Summary & Bottleneck

QueryParameterValue and Range perform several avoidable allocations during parameter construction, wire serialization (toValuePb and toTypePb), and wire deserialization (fromPb):

• QueryParameterValue.fromPb unconditionally allocates Map<String, QueryParameterType> parameterTypes = new HashMap<>() on every recursive call, including primitive scalars (INT64, STRING, BOOL, etc.) and ARRAY elements where parameterTypes is never used.
• When type == StandardSQLTypeName.STRUCT, QueryParameterValue.fromPb iterates typePb.getStructTypes() twice. The first pass allocates a HashMap and $F$ placeholder QueryParameterValue objects and passes them to valueBuilder.setStructTypes(structTypes). Immediately afterward, for every non-null struct, valueBuilder.setStructValues(structValues) calls setStructTypes(ImmutableMap.copyOf(structValues)), completely overwriting the map and all $F$ placeholder QueryParameterValue objects constructed in the first pass.
• QueryParameterValue.Builder.setStructValues calls ImmutableMap.copyOf(structValues) twice on the same input map instead of sharing a single ImmutableMap instance between structTypes and structValues.
• QueryParameterValue.array(T[], StandardSQLTypeName) allocates new ArrayList<>() with default capacity 10 instead of array.length.
• QueryParameterValue.toValuePb() and Range.getValues() call Range.getStart() and Range.getEnd() twice each. Every call to Range.getStart() or Range.getEnd() allocates a new FieldValue wrapper over getStartInner() or getEndInner().
• QueryParameterValue.validateTimestamp(String) allocates a regex Matcher for ISO8601_TIMESTAMP_HIGH_PRECISION_PATTERN (\.(\d{10,})(?:\D|$)) on every timestamp string parameter, even when the fractional second has 9 or fewer digits.

Changes

• In java-bigquery/google-cloud-bigquery/src/main/java/com/google/cloud/bigquery/QueryParameterValue.java:
  • Allocate parameterTypes only inside the non-null StandardSQLTypeName.STRUCT branch of fromPb, and only execute the placeholder structTypes loop when valuePb == null || valuePb.getStructValues() == null.
  • Pre-size collections in fromPb, toTypePb, and array(T[], StandardSQLTypeName).
  • Call ImmutableMap.copyOf(structValues) once in Builder.setStructValues and pass the resulting ImmutableMap to both setStructTypes and setStructValuesInner.
  • Read getRangeValues().getStartInner() and getEndInner() directly in toValuePb() instead of allocating 4 FieldValue wrappers via getStart() and getEnd().
  • Guard ISO8601_TIMESTAMP_HIGH_PRECISION_PATTERN.matcher(timestamp) in validateTimestamp with hasHighPrecisionFraction(timestamp) so timestamps with 9 or fewer fractional digits skip Matcher allocation.
• In java-bigquery/google-cloud-bigquery/src/main/java/com/google/cloud/bigquery/Range.java, read getStartInner() and getEndInner() directly in getValues().
• In java-bigquery/google-cloud-bigquery/src/test/java/com/google/cloud/bigquery/QueryParameterValueTest.java, assert that structValues and structTypes share the same ImmutableMap instance and verify fromPb on null STRUCT values.

Expected Impact

• Eliminates 1 HashMap allocation per scalar, array element, and struct field in QueryParameterValue.fromPb.
• Eliminates $F$ dead AutoValue_QueryParameterValue allocations, 1 HashMap, and 1 duplicate ImmutableMap.copyOf per non-null STRUCT parameter.
• Eliminates 4 FieldValue allocations per RANGE parameter during toValuePb() and Range.getValues().
• Eliminates regex Matcher allocation for all standard (up to nanosecond precision) timestamp string parameters.

Verification

• Ran unit tests in QueryParameterValueTest and RangeTest:

  mvn test -pl java-bigquery/google-cloud-bigquery -Dtest=QueryParameterValueTest,RangeTest -Dclirr.skip=true -Denforcer.skip=true -Dflatten.skip=true -Dcheckstyle.skip=true -Dfmt.skip=true